Dublin R - May 2016

Pat O’Reilly - AllState

“Anyone can fit an elephant”

Regularisation

L1, L2, ElasticNet

Island with gold reserves \(\rightarrow\) Predict deposits

Never a good thing

“That would be really cool!”

“You should definitely do that”

“So how are you thinking of generating the data…”

Basic Idea

Mine gold, make profit

Bid price, extraction costs, profit margins

Design Goals

Emphasise use of prediction, accuracy less important

Something for everyone

Vickrey auction: Highest bidder wins, pays second highest price

Easy to understand, code, fast to run

Harder to ‘game’

Starting capital

50 parcels of land, sequential auctions

\[ \text{Score} = \text{Start} - \text{Bids Paid} - \text{Extraction Costs} + \text{Sales} \]

Generating Data

Online datasets pointless

How do I generate data?

Make it up!

Hard to reverse engineer

No ‘true’ model

2D Gaussian Processes

\[ 150 \times 150 = 22,500 \; \text{land parcels} \]

Tested with simple models

Submissions

5,000 parcels in training set

Team submits bids for each parcel

Sample Submission

## parcel_id,bid_amount
## 515,734503.19
## 914,883127.20
## 1538,56130.52

Simple Shiny app

CSV upload with time stamp

Manual submissions - emails

12 submissions

Results Night

Facebook, mid-September 2016

7-8 teams

Most teams single person

Speakers described approach (5-10 mins)

Clustering, Cersei Lannister, Qlik

Mini-competitions

Hospital

Final Scores

• 3rd Place (54,468,017)
• 2nd Place (55,418,322)
• 1st Place (96,020,955)

Winning Approach

Unsophisticated model

Use variables to predict unknown variables to then predict gold

Focus on most profitable parcels

Overspend on auctions

What I Learned

I am bad at web code

It was a lot of fun

Definitely want to do it again

On that note …

El Dorado 2

I am Henry the VIIIth I am

(Rough) Schedule

Tonight: Expressions of Interest

http://eepurl.com/dMFjPc

December: Competition Launch

Jan/Feb: Competition Close and Results Night

Any Questions?

mickcooney@gmail.com